Neuromorphic computing was first proposed by Carver Mead in the late 20th century. Compared with traditional von Neumann computing, it can handle the explosive growth of data demand with its advantages of low energy consumption, serial parallel processing, and high energy efficiency [1–3]. At present, neuromorphic computing has shown great potential in many fields such as speech recognition [4] and image recognition [5]. Like biological synapses, synaptic devices are the fundamental components of mimetic neuromorphic computing [6,7]. At present, synaptic devices mainly include electrical and photoelectric synapse devices. Compared with electrical synapse devices, photoelectric synaptic devices [8–10] have been considered as a more suitable means to build an intelligent sensory system integrating sensory, memory, and computing functions, and they are expected to promote the development of retinal bionic technology [11]. In recent years, many photoelectric synapse devices have been investigated and improved their properties by using different materials, including oxides [12,13], perovskite [14,15], two-dimensional semiconductor materials MoS2 [16–18] or HfS2 [19], etc.

Recently, we also utilized ion-gel side-gated graphene to realize a photoelectric synaptic transistor with basic electrical synaptic functions and photoelectric collaborative learning [20]. However, low generation efficiency of photon-induced carriers in graphene makes it difficult to produce the photo postsynaptic current under short-time light pulse stimuli. Thus, it is necessary to introduce new material with high light response or construct new structures into the synaptic graphene transistor.

In recent years, CsPbBr3 has been considered as an attractive and promising optical material, but it has low charge carrier transfer performance [21,22]. Therefore, it must be considered to construct a heterojunction structure with a graphene or carbon nanotube [23–25]. The heterojunction structure can effectively superpose the advantages of CsPbBr3 and the graphene or carbon nanotube, thus prolonging their lifetime and effectively improving light sensitivity and responsivity [23–26]. Zhu et al. combined carbon nanotubes and perovskite quantum dots for an efficient neuromorphic vision system with high responsivity of 5.1×107A/W [25]. Pradhan et al. used a graphene-PQD (G-PQD) superstructure prepared by growing PQDs directly from a graphene lattice to mimic key bioequivalence characteristics and possess unique light enhancement and electrical habituation functions [26]. However, they all do not realize the negative photo-synaptic response yet, either by employing PQDs/carbon nanotube structures [25] or graphene-PQD structures [23,24,26]. But in human retina, cone cells and bipolar cells (positive and negative photoresponse) are crucial neurons. Specifically, the bipolar cells can respond to a light stimulus in excitatory and inhibitory manners. The main challenge in the application of the CsPbBr3/graphene structure for mimicking human retinotopic vision is to achieve a negative photoresponse characteristic.

Herein, to realize positive and negative photoresponse on the CsPbBr3/graphene heterojunction device, ion-gel as a gate dielectric material is utilized on the CsPbBr3/graphene heterojunction and graphene for controlling the charge transfer. CsPbBr3 is utilized as the light absorber layer and graphene as the carrier transmission layers. Depending on the pinning effect of CsPbBr3 nanocrystal, different photo-generated carriers can be transferred from CsPbBr3 to graphene by ionic-grated tuning of the Fermi level of graphene. It paves a way to develop human retinotopic vision neuromorphic systems for CsPbBr3/graphene photoelectric synapses with sensory, storage, and computing capabilities.

Figure 1(a) shows the structure of synapses that is composed of the presynaptic membrane, postsynaptic membrane, and neurotransmitter vesicles. When the pre-synapse is stimulated by external light or electrical pulse signals, potential changes occur within the presynaptic membrane, and then neurotransmitters are released through the vesicles into the synaptic cleft, resulting in a change in the postsynaptic membrane potential and then an excitatory or inhibitory response. Figure 1(b) shows the schematic illustration of our fabricated CsPbBr3/graphene heterojunction synaptic field effect transistor (FET) with an ion-gel side-gated structure design. Using PI as the substrate material, planar gate-drain-source Au electrodes [Pd (15 nm)/Au (50 nm)] were fabricated on PI by sputtering and lithography technology. PMMA-assisted single-layer graphene on copper foil purchased from XFNANO was standard-wet transferred to the source, drain electrodes, and PI substrate, and then the PMMA was removed using acetone. The graphene channel was patterned using photolithography and etched in oxygen plasma. After transferring CsPbBr3 nanocrystal solution (purchased from Mesolight) to part of graphene film surface by using dot-coating technology, the prepared ion gel solution was also dropped on perovskite, graphene, and the gate electrode for protecting the perovskite and graphene. Finally, the volatile solvent of the ion-gel solution was removed with a hot plate at 100°C. Part of the graphene without the CsPbBr3 nanocrystal is the guarantee of direct contact between the ion-gel and graphene. The CsPbBr3 and graphene surfaces were further characterized by an atomic force microscope (AFM), as seen in Figs. 1(c) and 1(d), where the CsPbBr3 nanocrystal with ∼30nm size is uniformly dispersed on the graphene film.

We have tested the I−V output curve of the device under the read voltage (VDS) of 0.05 V. As shown in Fig. 1(e), the gate voltage (VGS) is scanned from −3.5V to 3.5 V and then reversed. There, hysteresis windows have been observed. In addition, it can be seen that the Dirac point voltage of the device is around 1.5 V, which means when the gate voltage is less than 1.5 V, the device exhibits p-type and the charge carriers are holes. To investigate the effect of 450 nm light on its I−V curve, we tested the gate voltage scan from −3.5V to 3.5 V by irradiating 450 nm blue light with various powers (0, 10, 20, 30, and 40 mW) when VDS was 0.05 V. The results in Fig. 1(f) demonstrate that Dirac dot voltage gradually shifts to the right with the increase of light power, thus leading to higher currents. This indicates that the light causes charge transfer between CsPbBr3 nanocrystal and graphene.

Changes of synaptic weights are the basis of synaptic learning and memory processes. In our synaptic device, PSCs response changes (ΔPSCs) represent changes of synaptic weights. It is important to obtain different PSCs responses by applying different light/electrical stimuli pulses to analyze the factors affecting the change of the synaptic weight. In Fig. 2(a), ΔPSCs at VDS 0.05 V are obtained by 450 nm light stimuli with different powers. With the increase of the light power, ΔPSC increases from ∼0.27 to ∼0.92μA. In addition, the effect of the pulse width of 450 nm light on ΔPSC has also been analyzed, as shown in Fig. 2(b). With the increase of pulse width from 1 to 20 s at a fixed power of 30 mW, ΔPSC at 0.05 V read voltage also increases from ∼0.14 to ∼1.11μA. The increase trend of ΔPSC gradually shows a saturation state. This is because under short-time light stimuli the carrier’s generation rate is larger than the recombination rate, thus leading to rapid growth of ΔPSC, while as the light pulse width increases, the increased carrier’s recombination rate is close to the generation rate, resulting in the saturation of the channel photocurrent. In a word, the microampere level of photocurrent response about our device is more consistent with light-sensitive biosynaptic device applications. Subsequently, in Fig. 2(c), the ΔPSC response from ∼−34.6 to ∼−44.3μA is stimulated by the 0.8 V gate pulse with different pulse widths from 200 ms to 1 s. By improving the gate-voltage of the 500 ms electric pulse from 0.02 to 0.8 V, ΔPSC response changes from ∼−0.6 to ∼−41.5μA in Fig. 2(d). The results indicate that a larger pulse width and higher pulse voltage can greatly influence ΔPSC response. That is because many ions can be accumulated at two interfaces of the CsPbBr3/graphene channel and gate electrode by the gate voltage to form the electric double layers.

As one of important manifestations of short-term plasticity (STP), paired pulse facilitation (PPF) represents the real-time information processing of synapses. When the interspike interval (Δt) between paired light pulses is shorter than the recombination time of photo-carriers, the photocurrent triggered by the first light pulse cannot return to its original state before the next, resulting in the PPF effect. As shown in Fig. 2(e), the PPF index, defined as (A2/A1)×100%, is about ∼165% triggered by paired light pulses (2 s, 30 mW, Δt=100ms). The PPF index has a negative exponential relationship with the interspike interval Δt. Its function expression [27] is as follows: PPF index=C1exp(−Δtτ1)+C2exp(−Δtτ2)+C0,(1)where C0 is a constant, C1 and C2 are the initial facilitation degrees of the two phases, and τ1 and τ2 represent the characteristic relaxation time of the fast phase and the slow phase, respectively. As shown in Fig. 2(f), τ1 and τ2 are ∼1.6 and ∼49.7ms, respectively, and C1 and C2 are ∼31.3 and ∼132.9, respectively. τ2 is an order of magnitude larger than τ1, which is comparable to the time scale of biological synapses. Like the PPF effect by light stimuli, when Δt is less than the ion relaxation time, the ions triggered by the first pulse cannot return to their original state before the next, resulting in the electrical PPF effect of Figs. 2(g) and 2(h). The IPSC responses are triggered by paired 1 s and 1.5 V electrical pulses (VDS=0.05V, Δt=10ms), where the PPF index is about 117%. By changing the interspike interval of such paired electrical pulses from 1 to 2500 ms, the PPF value is changed from ∼117% to ∼106, which is comparable to the time scale in biological synapses.

Figure 3(a) shows the positive photocurrent response by changing the gate bias voltage VGS at 0 V and −1V, where the same light (40 mW, 3 s) stimulates. The ΔPSC value ∼0.95μA at −1 V gate bias is higher than that of 0 V gate bias (∼0.75μA). Without gate bias and light stimulation, the Fermi level of p-type graphene is slightly higher than that of CsPbBr3 nanocrystal because of their work function, so a potential barrier is formed, and their energy band bends upward in Fig. 3(b). When the light is applied, photogenerated electrons are excited to the conduction band of CsPbBr3, and under the built-in electric field photogenerated holes will be separated and transferred to the graphene. Parts of photogenerated holes participate in channel conduction, resulting in a positive photocurrent response, as shown in Fig. 3(a). An increased positive photocurrent response has also been observed under −1V gate bias. That is because applying −1V gate bias on the ion-gel enables deep p-type doping in the graphene and then down-shifts its Fermi level, which results in the decrease of the heterojunction potential barrier. This low potential barrier could lead to more photogenerated holes to transfer on the graphene, thus leading to an increased photocurrent response.

Subsequently, we further changed the gate bias voltage to −2.2V, and Fig. 3(c) shows the negative photocurrent response of about −0.5μA under the same light stimuli (40 mW, 3 s). This indicates that the photogenerated carriers in the CsPbBr3 cause the channel current to decrease. Due to the Fermi-level pinning effect of the CsPbBr3 [28], the energy band of the heterojunction changes from upward bending to downward bending [29], as shown in Fig. 3(d). That is because the ion-gel in direct contact with part of the graphene makes it easy to tune the Femi level of graphene. With the same light (40 mW, 3 s) stimuli, the photogenerated holes are held in the valance band of the CsPbBr3, while the photogenerated electrons are transferred from the conduction band to the graphene. These photogenerated electrons from the CsPbBr3 rapidly recombine with the charge-holes in the graphene, resulting in the reduction of the actual charge hole concentration. The deeper p-type doping in the graphene with −2.5V gate bias can further increase the potential barrier of CsPbBr3/graphene, thus leading to more photogenerated electrons being transferred into the graphene to form inhibitory photocurrents of ∼−0.63μA. The higher the negative gate bias is, the larger the inhibitory photocurrent becomes. Thus, our synaptic device exhibits positive and negative photoresponses dependent on the gate voltage of the ion-gel, resembling the biological characteristics of photoreceptors and bipolar cells.

In a word, we obtained the bipolar photocurrent response by changing gate bias, which solves the problem of synaptic devices having difficulty obtaining inhibitory photocurrent. Subsequently, the bipolar photocurrent response on the STP that is expanded to long-term plasticity in our device is obtained by increasing the number of consecutive light spike stimuli under different gate bias.

In Fig. 4(a), ΔPSC response is triggered by 40 successive light pulses (30 mW, 1 s) with various frequencies of 250, 400, 500, and 800 mHz. Higher-frequency consecutive light pulses are applied to generate larger long-term potentiation (LTP) at VDS=0.05V and VGS=0V. To detect the trend of high-pass filtering on the light signals, we extracted the gain value of the ΔPSC (A40) caused by the 40th pulse compared with that of the 1st pulse versus frequency, as shown in Fig. 4(b). The gain value shows a significant monotonic increase with increasing stimulation frequency, up to 5 at 800 mHz. To confirm the long-term inhibitory photocurrent characteristics of our device, a −2.5V gate voltage and VDS=0.05V are applied on the device, and then 1 min successive light pulses with various frequencies of 250, 400, and 500 mHz in Fig. 4(c) are utilized as stimulus. With the increasing light pulse frequency, the long-term depression (LTD) effect is more obvious. Similarly, we extract the value of the ΔPSC (B40) caused by the 40th light pulse compared with that of the 1st light pulse versus frequency, as shown in Fig. 4(d). Correspondingly, a gain value of 6 is obtained at the frequency of 500 mHz. These results also show that the device not only presents the LTP and LTD on the light pulse by controlling the gate bias, but also possesses a spike rate dependent plasticity (SRDP) on light pulses, where the plasticity weight is dependent on the frequency of the light spikes.

Figure 4(e) demonstrates the optical memory and electrical erasure effect. The LTP effect is related to the learning memory state by applying 30 consecutive optical spikes at the frequency 816 mHz (30 mW, 500 ms). Subsequently, 30 consecutive electrical spikes at the frequency of 1 Hz (10 mV, 500 ms) are applied on the device to obtain LTD for forgetting. This indicates that our device can be utilized as a programmable memory storage device by applying optical memory and electrical erasure. In Fig. 4(d), only applying successive electrical pulses at the frequency of 1 Hz (4 mV, 500 ms), the postsynaptic current exhibited an LTD response related to the oblivious state of the device. Subsequently, the same successive electrical pulses are continued, and 450 nm light illumination with 30 mW turns on, where the PSC exhibits an LTP response related to the memory recovery state of the device. Thus, the introduction of light illumination can shift the electrical synaptic response of the device from LTD to LTP. That facilitates the transition from the oblivious state to the memory learning state, which can enable the device to obtain a light-assisted re-learning function in the neuromorphic computing networks.

In addition, Morse code communication with 20 words per minute and “XOR” logic operation of our device have been realized in our device, as seen in Fig. 5. In Fig. 5(a), Morse code can be transmitted optically through our device, and different lighting stimuli durations can correspond to dot signals and dash signals in Morse code. Figure 5(b) demonstrates that the transmission and decoding of Morse code “BUPT” are performed by the ΔPSC response on such synaptic transistors with light stimuli of different durations. 450 nm light can be chosen as 30 mW stimulating light with the light pulse width 2 s for the dash signal (“−”), 0.5 s for the dot signal (“·”), and no light in 1 s between the dashed and dotted signals. Corresponding to the four letters of “BUPT,” we compiled 450 nm light pulses with one dash and three dots as “B,” two dots and one dash as “U,” one dot and two dashes and one dot as “P,” and one dash as “T,” respectively. When we applied them to our synaptic device, its ΔPSC response carried the Morse code information that had been obtained. The Morse decoding is obtained by measuring the numbers and values of ΔPSC response peaks. This is because the ΔPSC response is related to the duration of light pulses as well as the interval between two adjoining light pulses, due to short-term plasticity by one- and two-pulse light stimulus. This indicates that our device has some decoding capability and can be applied in the future fields of artificial intelligence signal sensing coding and decoding.

Furthermore, we also demonstrated the “XOR” logic operation of our device. The schematic diagram and status table of the “XOR” logic operation are shown in Fig. 5(c), where the light signal and electrical signal are input signals input from the top of the device and the gate electrode, respectively. A 450 nm light pulse with 30 mW light power and a pulse width of 1 s or an electrical pulse with an amplitude of 4 mV and a pulse width of 2 s can be considered as a signal “1,” and no light or electrical pulse is as the state “0.” We defined ΔPSC in the range of −0.1μA to 0.1 μA as output state “0” and ΔPSC larger than 0.1 μA or less than −0.1μA as output state “1.” Its output state of our device can be judged by observing the post-synaptic ΔPSC response. As shown in Fig. 5(d), when the logic input state is “10,” ΔPSC is obviously greater than 0.1 μA, and the device responds to the output state as “1”; when the logic input state is “01,” ΔPSC is significantly less than −0.1μA, and the device response output state is “1”; when the logic input state is “11,” there is a positive and negative ΔPSC response ranged between −0.1μA and 0.1 μA, and the device output state is “0”; when the logic input state is “00,” there is no response, and the device output state can be judged as “0.” Compared to the truth table, the device implements the “XOR” logic operation. In Fig. 5(d), the positive and negative ΔPSC response has happened in our device with the simultaneous stimuli of the optical signal and electrical signal. That is because of their different mechanisms of photo-synaptic currents and electrical-synaptic currents; as mentioned above, the photo-synaptic current is due to the photogenerated carrier’s separation and transfer, while the electrical synaptic current is due to the interfacial carrier trapping and the ionic relaxation effect. Moreover, the two mechanisms have different relaxation times for generating photo-synaptic and electrical synaptic currents. The relaxation time of generating the photo-synaptic current is significantly shorter than that of the electrical, and thus a more pronounced bipolar current jump occurs. The experiment demonstrates the hybrid photoelectric operation capability of the device and its potential application in the field of logic operations.

To further explore the typical perceptual learning capability of our device, we measure the pulse photocurrent of our synaptic device under different power densities and then simulate a 28×28 photocurrent array of the optical image of “7,” which was selected from the learning numbers in the Modified National Institute of Standards and Technology (MNIST) data set with 60,000 training images and 10,000 test images. Herein random noises are added to the input MNIST image related to adding light powers. We build a three-layer artificial neural network for handwritten font recognition with a multilayer perceptron (MLP), whose corresponding simulator utilizes 784 input neurons, 100 hidden neurons, and 10 output neurons, each corresponding to 10 numbers (0–9), as shown in Fig. 6(a). The neurons between different layers are connected by synapses, and the strength of the connection represents the synaptic weight, which updates, and its network trained by using back-propagation methods. The pixel matrix of the image is transformed into one dimension and transmitted to the input layer neurons. In Fig. 6(b), we change the initial state of the device by changing the light intensity, and then train the neural network. Under 30 cycles, the recognition accuracy of the neural network reached saturation, and the accuracy reached about 95% at the light current intensity of 40 mW. This indicates that our device has a very excellent achievement. The output of digital “7” under different optical powers of the device after training is shown in Fig. 6(c). After 30 training cycles, the accuracy of device training under different optical power input is different, and the output image clarity is also different. It can be seen from the figure that with the decrease of optical power, the definition of the picture is also decreasing.

A side ionic-gated CsPbBr3/graphene heterojunction photoelectric transistor has been fabricated and investigated. Basic biological synaptic plasticity like excitatory and inhibitory PSC, PPF, LTP, and LTD has been successfully mimicked under different optical or electrical pulse stimuli. Most importantly, based on the ionic-gate-tunable Fermi level of the graphene and the pinning effect of CsPbBr3, the bipolar photocurrent responses related to STP and LTP have been obtained by adjusting the negative gate bias from 0 to −2.5V. Moreover, the application of the device in Morse code, logic operation, the optical memory and electrical erasure effect, and light-assisted re-learning function has also been simulated. Furthermore, the handwritten font recognition accuracy of a neural network related to excellent optical synaptic behaviors was ∼95% with 30 training epochs on the MNIST dataset. These results provide new directions for the development of optoelectronic synaptic devices and future integrated intelligent sensory neuromorphic computing systems.